Below is my final report for the research project I collaborated on with Professor Mai Vu and PhD student Beier Li during the course of my master’s program. In brief, the topic was the robustness of different DNN and CNN architectures in response to noise added to different layers, as a follow up to this paper that proposed that initial layers matter more, and that noise added to initial layer degrade the overall model performance more than noise added to deeper layers. The findings were partly, but not fully reproducible. The GitHub repo for the experiments we ran can be found here.